Defining Transition Rules with Reinforcement Learning for Modeling Land Cover Change

Abstract

Spatio-temporal modeling provides the opportunity to simulate geographic processes of land use and land cover change (LUCC) by integrating geographic information systems (GIS) with various machine learning approaches to computing. Contemporary models are often developed using a training dataset to define a set of probabilistic transition rules that govern how a landscape changes over time. However, the use of training datasets can be problematic for spatio-temporal modeling, as they can limit the ability to incorporate system complexity and hinder the transferability of the model to different datasets. The purpose of this study is to evaluate a machine learning approach called reinforcement learning (RL) for defining transition rules for GIS-based models of land cover change due to natural resource extraction. Specifically, RL is evaluated based on its potential for constructing models independent of training datasets that can handle different levels of complexity and be transferred across different spatial extents. An RL model for Land Cover Change (RL-LCC) is developed for considering economic and ecological goals involved in natural resource management, and implemented using a hypothetical forest management scenario. Simulation results reveal that agents in the RL-LCC model are able to develop transition rules from their experience in their landscape in a variety of simulation scenarios that allow them to achieve their goals. This study demonstrates the benefits of integrating RL and GIS in order to address important issues of space, time and complexity.